Will ≥20 Atlas Fellows be actively working on ambitious projects aimed at large-scale problems? (M2500 subsidy)
Standard
36
Ṁ2761
resolved Oct 4
Resolved as
50%

The Atlas Fellowship is a scholarship and in-person program that helps talented young people understand the world and change it. In 2022 and 2023, there was a combined total of approximately 230 fellows. The program has been discontinued, so there won't be further fellows in the future.

Example projects that qualify (if there is recent progress):

- https://bioteinresearch.ca/

- https://www.airbound.co/

- https://www.youtube.com/watch?v=1QY1L61TKx0

- https://www.lesswrong.com/posts/3ecs6duLmTfyra3Gp/some-lessons-learned-from-studying-indirect-object

- https://www.lesswrong.com/posts/5spBue2z2tw4JuDCx/steering-gpt-2-xl-by-adding-an-activation-vector

Example projects that I find very exciting, but are slightly below the threshold for the purpose of this market:

- Studying hard subjects in college in order to apply the skills learnt to global problems (e.g. AI safety)

- Bringing the Burkina Faso team to the International Math Olympiad for the first time

- Participating in https://arena.education/

- https://payatric.com/ (insufficient visible progress)

- Getting physical copies of Harry Potter and the Methods of Rationality listed on Amazon

This resolves subjectively according to my opinion. For a project to count, the Atlas Fellow needs to be working on it actively with a significant time commitment and visible recent progress at the time of resolution (1 Oct 2024); it is not sufficient for them to have previously worked on it but then completed or abandoned it.

I may place trades in this market. I promise to try really hard to resolve this according to my best judgment, rather than what maximizes my Mana income. If you would be more likely to trade in this market if I didn't, please let me know in the comments and I'll reconsider.

Get
Ṁ1,000
and
S1.00
Sort by:

Resolving this to 50%:

I think this the final number is around 20, and there are a few cases that are ambiguous enough that I think the overall resolution is kind of ambiguous. If you put a gun to my head and forced me to pick an option right now, flipping a coin would feel reasonable. If I invested a lot more time into figuring this out, I'd land somewhere between 16 and 28.

The market description states: "This resolves subjectively according to my opinion." Resolving to a percentage seems in line with that description.

If people somehow get super angry about this, I could instead ask the mods to cancel the market, but I think a 50% resolution is reasonable.

@JonasVollmer Seems reasonable. I would personally have resolved that to YES, on the grounds that you're likely to have missed at least one such project? But 50% seems defensible too

@NeelNanda The above conclusion was factoring in that I may have missed a few projects

@JonasVollmer I'm curious how counterfactual this result is. Would these people have done a similar number of ambitious projects without doing Atlas? Would they have done different ambitious projects?

I'm working on resolving this! If you're aware of any Atlas Fellows doing cool projects, please mention them in the thread :)

So far identified seven projects that meet the criteria

@JonasVollmer @eternalrecursion the Vienna Hypersphere!

@logaems Oh what's that?

(Currently at 13 projects)

@JonasVollmer A new theoretical-focused alignment group (@mudPi314 and a few other fellows are also involved). However, I just noticed the market closed, so this probably doesn't count, as it has just begun.

How are you going to find the information to count this?

predicted YES

Would there be interest in an identical market with a threshold of "≥ 10 Atlas Fellows"?

@JonasVollmer i think this is almost certainly yes, if u disagree a market might be good (:

@JonasVollmer I’m curious what safety research qualifies for this. Ie. does IOI count because it was an especially good mech interp paper, or would an average-quality interp paper also count? This changes my probability a lot

predicted YES

@Asher An average-quality interp paper would probably count, though it depends on the details.

I would count a research project if it broadly meets the following criteria:

- Aimed at a large-scale problem. This is much broader than AI safety, e.g. I'd count econ growth diagnostics, longevity research, Hepatitis E vaccines, cost-benefit analyses of drug review policies, private mercenary companies, HFC phase-outs, etc.

- Good ideas that could potentially push the frontier in a field, rather than just making minor iterations on existing research. So, an interpretability paper that's just a replication wouldn't count, unless it's targeting an especially influential paper and there are reasons to doubt it'll replicate.

- At least at the level of quality and insight that you might expect from an above-average MIT undergrad; lower-quality research would not count.

- I will give relatively little weight to meeting academic peer review standards, partly because I don't understand them very well, partly because I think they're overrated, partly because I don't think it makes sense to apply them to early-career individuals.

I recognize this is extremely subjective, so feel free to ask me for specific examples and I'll try to respond. Also if you think I shouldn't bet in this market due that, let me know.

predicted YES

@JonasVollmer Makes sense, thanks! Yea these seem like good criteria

Do you expect there to be more Atlas fellows?

predicted NO

@RichardAnderson No—see the following market:

Added Ṁ1k subsidy.

predicted NO

I'd say the current number is somewhere around 2–7, haven't made a strong effort to nail it down precisely.

predicted YES

@JonasVollmer will u make the strong effort near the closing date?

predicted YES

@uzpg I was hoping to make it more like a crowdsourced project, but yeah, I will be spending some time on it then!

predicted NO

Some people wondered if the projects need to be EA-related to count, or if profits need to be donated to EA causes to count. The answer is no; e.g. if an Atlas Fellow launches a company that will likely make them >$10m/y by solving a large-scale problem for their customers, or a research project that has a similar effect, that counts for the purposes of this market.

By what date do they need to be working on those problems, in order to count? The market close date?

predicted NO